SamaraAD

The Offical Code of CVPR2024: Adversarial Distillation Based on Slack Matching and Attribution Region Alignment

This repo is under construction.

Usage

Slack Matching

You can implement $L_{sm}$ by doing the following:

from spearman_soft_rank import spearman_soft_rank

y_student = student_model(x)
y_teacher = teacher_model(x)
loss = spearman_soft_rank(y_student, y_teacher, beta)

Attribution Region Alignment

You can easily get the class activation mapping for the model using this repo and then use MSE to compute $L_{align}$.